Cumby-Huizinga test for autocorrelation

[Alexandra TD]

Hello. I have some difficulties with the interpretation of Cumby-Huizinga test for autocorrelation. I found online the presentation "a general approach to testing for autocorrelation" written by professor Christopher F Baum and Mark Schaffer.
Regarding this presentation, and looking for the examples at page 20 and 35 , how should I interpret the results for actest here ?
I have been looking online regarding this but I dont seem to find anything.
Thank you,
Alexandra


http://fmwww.bc.edu/EC-C/S2014/823/UKSUG2013.pdf
or
http://www.stata.com/meeting/new-orl...ola13-baum.pdf

[Eric de Souza]

A new user-written command (by Christopher Baum) called actest is now available.
The simplest form is actest, lags(#), where # stands for the lag order. For example, actest, lags(4).
actest performs two kinds of tests.
The first is similar to the bgodfrey test for each of the lags up till # (left hand column).
The second is that the autocorrelation at each lag order is zero, allowing for non-zero autocorrelations at lower lag orders (right hand column).
actest, lags(4)
Cumby-Huizinga test for autocorrelation (Breusch-Godfrey)
H0: variable is MA process up to order q
HA: serial correlation present at specified lags >q
-----------------------------------------------------------------------------
H0: q=0 (serially uncorrelated) | H0: q=specified lag-1
HA: s.c. present at range specified | HA: s.c. present at lag specified
-----------------------------------------+-----------------------------------
lags | chi2 df p-val | lag | chi2 df p-val
-----------+-----------------------------+-----+-----------------------------
1 - 1 | 93.148 1 0.0000 | 1 | 93.148 1 0.0000
1 - 2 | 93.276 2 0.0000 | 2 | 28.883 1 0.0000
1 - 3 | 93.276 3 0.0000 | 3 | 15.921 1 0.0001
1 - 4 | 93.330 4 0.0000 | 4 | 10.636 1 0.0011
-----------------------------------------------------------------------------
Test allows predetermined regressors/instruments
Test requires conditional homoskedasticity

If we only wish to test that the third-order autocorrelation is zero, the command would be:
. actest, lags(3 3)

Cumby-Huizinga test for autocorrelation (Breusch-Godfrey)
H0: variable is MA process up to order q
HA: serial correlation present at specified lags >q
-----------------------------------------------------------------------------
H0: variable is MA(q), q=2 | H0: q=specified lag-1
HA: s.c. present at range specified | HA: s.c. present at lag specified
-----------------------------------------+-----------------------------------
lags | chi2 df p-val | lag | chi2 df p-val
-----------+-----------------------------+-----+-----------------------------
3 - 3 | 15.921 1 0.0001 | 3 | 15.921 1 0.0001
-----------------------------------------------------------------------------
Test allows predetermined regressors/instruments
Test requires conditional homoskedasticity

If we wish to test that the second- and third-order autocorrelations are zero, the command would be:
. actest, lags(2 3)

I will not be following up on this for the moment

[Alexandra TD]

Dear Eric,
Thank you.

From what I understand , this means for example that regarding page 22 in the presentation, the conclusion would be that we have an MA process up to lag 4 .
At page 35 that the lags are higher than 4 , under conditions of heterosk.
What about page 40 ? Because in the left column it means that we have serial correlation at range 1-4 and in the right column for the two first lags we have serial correlation, for lag 3 we dont and lag 4 we do ? How do I interpret this one ? How will the probability of 0.2816 affects my interpretation ?
At page 46 I have two p-values in the right column>0.5 . , would this mean that I have serial correlation at lag 1, 2 and 4 ?
Page 56: right column :I have serial correlation at lag 1-8 ( up to lag 8, as tested) and in the right serial correlation up to lag 4 ?


I clearly am making a mistake in understanding these tables, could you help me see where ?
Thank you,
Alexandra

[Eric de Souza]

I am too busy with exam corrections and class preparation for the moment. But your page numbers are odd: there are only 44 slides in the presentations.

[Mark Schaffer]

Alexandra - I think the confusion over page vs. slide numbers is that there are only 44 slides, but slides appear multiple times because of highlighting etc. so there are 103 pages in total. I think you are asking about slide 18, which appears on pp. 39-41 of the PDF.

You ask: "Because in the left column it means that we have serial correlation at range 1-4 and in the right column for the two first lags we have serial correlation, for lag 3 we dont and lag 4 we do?"

Yes, that's basically it.

The tests reported on the LHS are standard tests for the presence of autocorrelation. The null is that there is no autocorrelation and the alternative is that there is autocorrelation in the specified range. For example, the test stat reported for "lags 1-3" will reject if there is serial correlation anywhere in lags 1-3, but has no power to detect autocorrelation at higher lags.

The tests reported on the RHS are tests for the presence of autocorrelation at the specified lag. For example, the test stat reported for "lag 3" will reject if there is serial correlation at lag 3 but is not designed to detect autocorrelation at other lags.

The example on slide 20 is a bit better because the null for the RHS test is a bit more sensible (H0: there may be autocorrelation up to lag 2 but not at lag 3, and HA: there is autocorrelation at lag 3) but the picture is the same.

The interpretation of the tests on the LHS is that you reject H0: no autocorrelation because however far back you look (lag 1 only, lags 1-2, lags 1-3, lags 1-4) you find autocorrelation.

The interpetation of the tests on the RHS is that there is evidence of autocorrelation at lag 1, lag 2, and lag 4, but not at lag 3.

[Alexandra TD]

Dear Mark, Thank you.
I am checking these because I am trying to decide the number of lags to be considered in newey west regression for panel data. If I have autocorr at lag 1, 2 4, but not 3, what is the number of lags I will consider in newey west? 4 ?
Thank you again,
Alexandra

[Mark Schaffer]

See my comment here: http://www.statalist.org/forums/foru...-fixed-effects